From Human Explanation to Model Interpretability: A Framework Based on Weight of Evidence